The Line-Up
// one single track.

Apple's Trusted Execution Monitor (TXM) is one of the most security-critical components of the modern iOS and macOS architecture: running alongside SPTM at a privilege level above XNU, it is the arbiter of code-signing and trust decisions for the entire system. It is also small, closed, undocumented, and painful to analyze — exactly the kind of target where a single researcher's time doesn't scale.

This talk is about changing that ratio by wiring large language models directly into a vulnerability-research pipeline. We begin with binary diffing across firmware releases to find what changed and where to look, then progressively hand more of the loop to the machine. We rehost TXM in userland so it can be executed and instrumented off-device, and let an AI agent write the fuzzing harnesses that drive it. From there we move TXM into a hypervisor-backed environment for higher-fidelity execution, and finally let the agent decompile TXM and hunt for vulnerabilities directly in the reconstructed code.

Along the way we'll be honest about what works, what breaks, where the AI confidently lies to you, and where it genuinely outperforms a human reviewer — leaving attendees with a realistic picture of agentic vulnerability research against a hard, real-world target, and a blueprint for pointing the same techniques at their own.